betaFeedback

WorksPhysics and Astronomy

Physics-informed neural networks (PINNs) for kinetic shock waves: simultaneous accuracy of lower- and higher-order moments, with and without moment supervision

Abstract

We investigate physics-informed neural networks (PINNs) for one-dimensional kinetic shock waves governed by the Bhatnagar–Gross–Krook (BGK) and Shakhov collision models, with and without moment supervision. The trace-free third-order and the fourth-order moments enter the fluxes of stress and heat flux through a shock; the aim, to our knowledge not previously examined for kinetic PINNs, is simultaneous accuracy of these and the lower-order moments of one predicted distribution. Data-assisted calculations, including anchor and closure studies, combine kinetic residuals with moment observations. For the BGK shock tube, a positive eight-moment representation whose spatial envelope follows the far-field decay of the initial condition gives heat-flux, stress and higher-moment discrepancies of 0.33%0.33\%, 0.67%0.67\% and 1.81.8–2.0%2.0\% against the 6400-cell discrete-velocity method (DVM) reference supplying its observations; a five-field representation leaves the latter uncontrolled. Equation-trained characteristic-integral PINNs use analytic initial and incoming-boundary distributions and their own predicted moments. Their BGK and Shakhov shock-tube heat-flux and stress discrepancies are 0.41%0.41\% and 0.36%0.36\% against 3200-cell references; against the 6400-cell reference the BGK values are 0.21%0.21\% and 0.18%0.18\%, and 0.40.4–0.9%0.9\% for the higher moments. Stationary neural continuation covers Shakhov Mach 2–5 and BGK Mach 6–7. Characteristic correction reduces the Mach-7 heat-flux and stress discrepancies from 3.37%3.37\% and 1.34%1.34\% to 0.42%0.42\% and 0.43%0.43\%; at Mach~8 they are 0.62%0.62\% and 0.79%0.79\%, and at Mach~12, across references, 0.400.40–1.03%1.03\% and 0.410.41–2.65%2.65\%. Three independent Mach-2 initialisations assess repeatability. It is established that lower moments do not determine higher ones; the contribution is to quantify which moments each information regime recovers.

Introduction

The structure of a rarefied shock is determined by a molecular velocity distribution. Density, velocity and temperature describe its bulk behaviour, but stress, heat flux and higher moments resolve different parts of its nonequilibrium structure. These quantities matter in rarefied and high-speed gas dynamics, where local equilibrium closures can become inaccurate [4, 36, 3, 17]. A numerical method must therefore be assessed through the moments needed by the application, rather than through agreement in a few bulk fields alone.

The Bhatnagar–Gross–Krook (BGK) model [2] replaces the Boltzmann collision integral with relaxation towards a local Maxwellian. It preserves the nonlinear coupling between a distribution and its moments, but gives unit Prandtl number. The Shakhov model [34] introduces a heat-flux correction to recover the monatomic value 2/3. Differences between collision models are visible in shock structure and higher moments [9] Shoja-Sani et al., 2021]. The conservative discrete-velocity method (DVM) and related kinetic schemes provide numerical references [27, 10, 8, 14, 45, 39].

Physics-informed neural networks (PINNs) approximate solutions by minimising equations and boundary residuals [30]. The PINN–BGK formulation has been applied to forward and inverse rarefied flows [25]. Separable neural representations address the cost of phase-space approximation [29], and velocity-weighted losses control the errors of the conserved moments [20]. Asymptotic-preserving, micro–macro and structure-preserving neural methods incorporate further kinetic structure [15, 16, 23, 5]. Related work concerns learned collision models and neural descriptions of rarefied flows [1, 33 a]. These approaches motivate two related questions: which moments become identifiable when selected observations are supplied, and what transport and higher-moment accuracy can be obtained from the kinetic equations alone?

The higher moments of the distribution, the trace-free third-order moment and the fourth-order moments, have a direct role in gas dynamics. In the balance laws for stress and heat flux obtained from the kinetic equation, they form the parts of the fluxes of these quantities that the lower moments do not determine. They vanish for the distributions underlying the Grad and, for the relaxation models used here, the Navier–Stokes–Fourier descriptions, so that their values inside a shock measure transport that those descriptions omit. In Grad’s and the regularised 13-moment equations they are the quantities whose closure determines the predicted stress and heat flux, and in the 26-moment equations they are variables [11, 38, 41, 37, 40, 12] (Levermore 1996). Their velocity kernels weight the tails of the distribution, which inside a shock are shaped by the overlap of the upstream and downstream populations [28], so that they amplify errors there. Agreement in lower moments does not by itself bound those projections. A finite-dimensional projection argument and velocity-space examples explain this distinction.

This work compares data-assisted and equation-trained PINNs for BGK and Shakhov shock tubes and stationary shocks. The neural calculations extend through Mach 7, and PINN-initialised characteristic BGK calculations cover Mach 7, 8 and 12. The comparison focuses on heat flux, stress and higher moments, together with the bulk fields and kinetic balances.

The data-assisted calculations determine what selected observations add to a finite kinetic representation. Moment anchors constrain transport quantities in the positive macro–micro shock-tube model, while a stationary closure head tests selective fourth-order recovery. For the base shock tube, the representation and the observations are extended to third- and fourth-order moments; the intermediate results show that these moments are limited first by the velocity kernels and then by the spatial envelope of the non-equilibrium amplitudes. These experiments are useful even when an equation-trained solution is available: they isolate which moment information improves a prediction and which quantities remain poorly determined by lower-order agreement. A constrained least-squares continuation couples the supplied moment information to separate field and physical-residual controls.

The equation-trained PINNs solve the forward problem using kinetic residuals, analytic initial and boundary states, and moments of their own predicted distributions. We consider direct kinetic-residual networks and characteristic-integral source networks. The stationary neural solutions use Shakhov collisions at Mach 2–5 and BGK collisions at Mach 6–7. For the stronger BGK shocks, characteristic correction updates macroscopic source coefficients with the neural weights fixed. This stage is a kinetic solve initialised by a PINN prediction. The calculations therefore examine both neural approximation and the improvement available from a subsequent transport solve.

The numerical inputs and reference-derived quantities are specified for each calculation. Comparisons use signed physical fields and collision-matched DVM references, with reference-grid sensitivity reported alongside neural discrepancies. All kinetic experiments are planar and one-dimensional.

The aim of this work is simultaneous accuracy of the lower- and higher-order moments: density, velocity, temperature, heat flux and stress, and the higher moments defined in § 2.1, integrated from the same predicted distribution and compared with the same reference. Because the higher moments enter the fluxes of stress and heat flux (§ 2.1), an approximate distribution whose lower moments are accurate but whose higher moments are not transports stress and heat flux incorrectly, so that its stress and heat flux are inconsistent with their own balance laws; for this reason the PINNs of this work are assessed on both, and the data-assisted base case is also trained on both. Neural kinetic solvers have been assessed mainly through their conserved fields or flow rates [25, 7, 16, 29, 20]. Stress and heat flux have been reported in a few cases: integrated from a distribution predicted by the network [46, 47], computed with a neural collision operator [43, 44, 33], regressed from reference data [32], or obtained as the outputs of a PINN that solves a closed moment system for a Mach-10 normal shock [18], discussed in § 11. Machine-learned moment closures evolve learned or higher moments as variables of a reduced moment system rather than integrating them from one predicted distribution [13, 6]. To our knowledge, whether the higher moments beyond stress and heat flux, integrated from the same predicted distribution, remain accurate alongside the lower-order fields has not been examined for neural kinetic solvers. Simultaneous accuracy is demanding for four reasons, examined in § 11. The higher-order kernels weight the velocity tails, where small distribution errors are amplified (§ 4; see also Xiao and Frank [44], Oh et al. [29]). As is well established, a loss that controls the lower moments leaves the higher projections unconstrained (Lemma 1). Once the higher moments are targeted, fitting them alone can degrade the lower ones unless both are controlled (§§§ 5.2 and § 5.5; Appendix § 13.§ 2). Finally, the higher moments of the reference carry larger discretisation errors, so that their discrepancies must be judged against reference refinement.

Section 2 defines the kinetic models and problems; section 3 describes the evaluation protocol. Section 4 explains moment sensitivity. Section 5 presents the data-assisted algorithms and experiments; section 6 develops the equation-trained shock-tube algorithm. Sections 8 and 9 describe the stationary algorithms, followed by their physical comparisons in section 10. Appendix § 13 provides the numerical configurations, self-scaling coefficients and independent-initialisation tests. Appendix § 14 summarises the reference comparisons and available refinement studies. Appendix § 15 presents the stage-resolved optimisation histories for the data-assisted and equation-trained calculations. The data-assisted schedules are detailed in Appendix § 13.§ 2 and tables 14 and 15.

Kinetic models and shock problems

Relaxation models

We consider a monatomic gas with one spatial coordinate and three molecular velocity components. All quantities are nondimensional. The gas constant, molecular mass, and reference temperature scale are absorbed into the nondimensional variables, and the translational degrees of freedom are three.

Both kinetic models used in this paper have the relaxation form

∂tf+vx∂xf=1τ(T[f]−f),v=(vx,vy,vz)∈R3,(1)\partial_t f + v_x \partial_x f = \frac{1}{\tau}\left(\mathcal{T}[f] - f\right), \qquad\boldsymbol{v} = (v_x,v_y,v_z) \in\mathbb{R}^3, \tag*{(1)}

where f(x,t,v)f(x,t,\boldsymbol{v}) is the molecular velocity distribution, xx is position and tt time; the relaxation target T[f]\mathcal{T}[f] is constructed from moments of ff and τ\tau is the relaxation time.

The BGK model takes T[f]=M[f]\mathcal{T}[f]=\mathcal{M}[f], where the local Maxwellian is defined by (2)

Mρ,u,T(v)=ρ(2πT)3/2exp⁡[−∣v−uex∣22T],(2)\mathcal{M}_{\rho,u,T}(\mathbf{v})=\frac{\rho}{(2\pi T)^{3/2}}\exp\left[-\frac{|\mathbf{v}-u\mathbf{e}_x|^2}{2T}\right], \tag*{(2)}

Here ex\mathbf{e}_x is the unit vector along the spatial coordinate. The density, streamwise velocity and temperature in Eqs. (3) to (5) are moments of ff:

ρ=∫R3f dv,(3)\rho=\int_{\mathbb{R}^3}f\,\mathrm{d}v, \tag*{(3)}
u=1ρ∫R3vxf dv,(4)u=\frac{1}{\rho}\int_{\mathbb{R}^3}v_xf\,\mathrm{d}v, \tag*{(4)}
T=13ρ∫R3∣v−uex∣2f dv.(5)T=\frac{1}{3\rho}\int_{\mathbb{R}^3}|\mathbf{v}-u\mathbf{e}_x|^2f\,\mathrm{d}v. \tag*{(5)}

For a monatomic gas, the corresponding equilibrium pressure is p=ρTp=\rho T and the ratio of specific heats is γ=5/3\gamma=5/3. The peculiar velocity is defined by (6)

c=v−uex.(6)\mathbf{c}=\mathbf{v}-u\mathbf{e}_x. \tag*{(6)}

The nonequilibrium heat flux and normal deviatoric stress used for validation are the central moments in Eqs. (7) and (8)

qx=12∫R3∣c∣2cxf dv,(7)q_x=\frac{1}{2}\int_{\mathbb{R}^3}|\mathbf{c}|^2c_xf\,\mathrm{d}v, \tag*{(7)}
σxx=∫R3(cx2−∣c∣23)f dv.(8)\sigma_{xx}=\int_{\mathbb{R}^3}\left(c_x^2-\frac{|\mathbf{c}|^2}{3}\right)f\,\mathrm{d}v. \tag*{(8)}

The sign convention in (8) follows the deviatoric second central moment used in the computations. These two quantities vanish for a local Maxwellian and therefore directly measure the nonequilibrium component of the shock layer. The third- and fourth-order nonequilibrium moments assessed in this work are the trace-free third moment mxxxm_{xxx}, the trace-free fourth-order moment RxxR_{xx} and the scalar fourth-order moment Δ\Delta [37],

mxxx=∫R3(cx3−35cx∣c∣2)f dv,(9)m_{xxx}=\int_{\mathbb{R}^3}\left(c_x^3-\frac{3}{5}c_x|\mathbf{c}|^2\right)f\,\mathrm{d}v, \tag*{(9)}
Rxx=∫R3∣c∣2(cx2−∣c∣23)f dv−7Tσxx,(10)R_{xx}=\int_{\mathbb{R}^3}|\mathbf{c}|^2\left(c_x^2-\frac{|\mathbf{c}|^2}{3}\right)f\,\mathrm{d}v-7T\sigma_{xx}, \tag*{(10)}
Δ=∫R3∣c∣4f dv−15ρT2,Sxx=Rxx+Δ3.(11)\Delta=\int_{\mathbb{R}^3}|\mathbf{c}|^4f\,\mathrm{d}v-15\rho T^2,\qquad S_{xx}=R_{xx}+\frac{\Delta}{3}. \tag*{(11)}

Like qxq_x and σxx\sigma_{xx}, these moments vanish for a local Maxwellian; SxxS_{xx} combines the two fourth-order moments. All higher-moment profiles in this paper use these full-distribution definitions, evaluated with the velocity quadrature of each calculation; the closure experiments of § 5.4 use the finite-grid subtractions stated there.

These moments have a direct gas-dynamic role. Multiplying (1) by cx2−∣c∣2/3c_x^2-|\mathbf{c}|^2/3 or by ∣c∣2cx/2|\mathbf{c}|^2c_x/2 and integrating over v\mathbf{v} gives the balance laws for σxx\sigma_{xx} and qxq_x. Besides the time derivatives, the convective fluxes uσxxu\sigma_{xx} and uqxuq_x, the collision terms and production terms proportional to ∂xu\partial_xu and to ∂tu+u∂xu\partial_tu+u\partial_xu, these balance laws contain the xx-derivatives of the fluxes relative to the mean flow,

∫R3cx(cx2−∣c∣23)f dv=mxxx+815qx,(12)\int_{\mathbb{R}^3}c_x\left(c_x^2-\frac{|\mathbf{c}|^2}{3}\right)f\,\mathrm{d}v=m_{xxx}+\frac{8}{15}q_x, \tag*{(12)}
12∫R3cx2∣c∣2f dv=12Sxx+72Tσxx+52ρT2.(13)\frac{1}{2}\int_{\mathbb{R}^3}c_x^2|\mathbf{c}|^2f\,\mathrm{d}v=\frac{1}{2}S_{xx}+\frac{7}{2}T\sigma_{xx}+\frac{5}{2}\rho T^2. \tag*{(13)}

In (12) the term 8qx15\frac{8q_x}{15} is fixed by the heat flux, and in (13) the terms 72Tσxx+52ρT2\frac{7}{2}T\sigma_{xx}+\frac{5}{2}\rho T^2 are fixed by the lower moments; the trace-free third moment mxxxm_{xxx} and the fourth-order combination SxxS_{xx} carry the remainder, that is, the part of the transport of stress and heat flux that the lower moments do not determine. The moment mxxxm_{xxx} also enters the heat-flux balance through the production term mxxx ∂xum_{xxx}\,\partial_xu. All three higher moments vanish for the Grad 13-moment distribution and, for the relaxation models considered here, for the first-order Chapman–Enskog distribution [37], on which the Grad and Navier–Stokes–Fourier descriptions respectively rest, so that their values inside a shock measure transport of stress and heat flux beyond those descriptions. In regularised 13-moment theory mxxxm_{xxx}, RxxR_{xx} and Δ\Delta are the quantities whose constitutive relations close the equations for stress and heat flux [38, 37], and in the 26-moment equations they are themselves variables [12]; kinetic solutions that give them accurately can therefore test such closures directly, rather than only through the stress and heat flux they predict. Physically, mxxxm_{xxx} is the trace-free part of the third central moment of cxc_x, ∫R3cx3f dv=mxxx+6qx5\int_{\mathbb{R}^3}c_x^3f\,\mathrm{d}v=m_{xxx}+\frac{6q_x}{5}, and Δ\Delta is the departure of the fourth moment of the peculiar speed ∣c∣\lvert c\rvert from its value for the Maxwellian with the same density and temperature, a measure of the kurtosis of the distribution. Inside a shock these moments arise from the overlap of the upstream and downstream populations [28]; for two overlapping Maxwellians, the separation of their mean velocities lowers Δ\Delta and the difference of their temperatures raises it, so that Δ\Delta can take either sign within the layer, as it does in the base shock tube (Figure 4). For these reasons the PINNs of this paper are assessed on the higher moments as well as on the lower ones.

Figure 4. Moment-supervised BGK shock tube at Kn=1\mathrm{Kn}=1 and tf=0.02t_f=0.02: eight signed fields of the five-field state, of the eight-moment state with the original envelopes, and of the eight-moment state with the common-tail envelope after the output-layer refinement, compared with the 6400-cell DVM reference. All displayed fields are integrated from the corresponding positive kinetic distribution. The eight-moment states use macroscopic, transport-moment and third- and fourth-order moment observations. In ρ\rho, uu, TT, qxq_x and σxx\sigma_{xx} the three states nearly coincide at the scale of the plot (Table 3 gives their discrepancies); the two eight-moment states differ most visibly in the upstream shoulder of mxxxm_{xxx}, RxxR_{xx} and Δ\Delta, which is enlarged in Figure 5. The original envelopes are GwG_w in (36) and GhG_h in (48); the common-tail envelope is (49), and the output-layer refinement is described in § 5.2.6.

Because the BGK target relaxes every nonequilibrium moment at the same rate 1/τ1/\tau, the model has unit Prandtl number. The Shakhov model [34] replaces the target by

References

  1. [1]Nicholas Daultry Ball, Jonathan F. MacArt, and Justin Sirignano. Online optimisation of machine learning collision models to accelerate direct molecular simulation of rarefied gas flows. Journal of Computational Physics, 549:114601, 2026. doi: 10.1016/j.jcp.2025.114601. URL https://doi.org/10.1016/j.jcp.2025.114601.
  2. [2]P. L. Bhatnagar, E. P. Gross, and M. Krook. A model for collision processes in gases. i. small amplitude processes in charged and neutral one-component systems. Physical Review, 94(3):511–525, 1954. doi: 10.1103/PhysRev.94.511. URL https://journals.aps.org/pr/abstract/10.1103/PhysRev.94.511.
  3. [3]Graeme A. Bird. Molecular Gas Dynamics and the Direct Simulation of Gas Flows. Clarendon Press, Oxford, 1994. URL https://academic.oup.com/book/54839.
  4. [4]Carlo Cercignani. The Boltzmann Equation and Its Applications. Springer, 1988. doi: 10.1007/978-1-4612-1039-9. URL https://link.springer.com/book/10.1007/978-1-4612-1039-9.DOI
  5. [5]Wei Chen, Giacomo Dimarco, and Lorenzo Pareschi. Structure and asymptotic preserving deep neural surrogates for uncertainty quantification in multiscale kinetic equations. arXiv preprint arXiv:2506.10636, 2025. doi: 10.48550/arXiv.2506.10636. URL https://arxiv.org/abs/2506.10636.
  6. [6]Andrew J. Christlieb, Mingchang Ding, Juntao Huang, and Nicholas A. Krupansky. Hyperbolic machine learning moment closures for the BGK equations. Multiscale Modeling & Simulation, 23(1):187–217, 2025. doi: 10.1137/24M1629377.DOI
  7. [7]Mario De Florio, Enrico Schiassi, Barry D. Ganapol, and Roberto Furfaro. Physics-informed neural networks for rarefied-gas dynamics: Thermal creep flow in the Bhatnagar–Gross–Krook approximation. Physics of Fluids, 33(4):047110, 2021. doi: 10.1063/5.0046181.DOI
  8. [8]Giacomo Dimarco and Lorenzo Pareschi. Numerical methods for kinetic equations. Acta Numerica, 23:369–520, 2014. doi: 10.1017/S0962492914000063. URL https://doi.org/10.1017/S0962492914000063.
  9. [9]Fei Fei, Haihong Liu, Zhaohui Liu, and Jun Zhang. A benchmark study of kinetic models for shock waves. AIAA Journal, 58(6):2596–2608, 2020. doi: 10.2514/1.J059029. URL https://doi.org/10.2514/1.J059029.
  10. [10]Francis Filbet and Shi Jin. A class of asymptotic-preserving schemes for kinetic equations and related problems with stiff sources. Journal of Computational Physics, 229(20):7625–7648, 2010. doi: 10.1016/j.jcp.2010.06.017. URL https://doi.org/10.1016/j.jcp.2010.06.017.
  11. [11]Harold Grad. On the kinetic theory of rarefied gases. Communications on Pure and Applied Mathematics, 2(4):331–407, 1949. doi: 10.1002/cpa.3160020403. URL https://doi.org/10.1002/cpa.3160020403.
  12. [12]Xiao-Jun Gu and David R. Emerson. A high-order moment approach for capturing non-equilibrium phenomena in the transition regime. Journal of Fluid Mechanics, 636:177–216, 2009. doi: 10.1017/S002211200900768X. URL https://doi.org/10.1017/S002211200900768X.
  13. [13]Jiequn Han, Chao Ma, Zheng Ma, and Weinan E. Uniformly accurate machine learning-based hydrodynamic models for kinetic equations. Proceedings of the National Academy of Sciences, 116(44):21983–21991, 2019. doi: 10.1073/pnas.1909854116.DOI
  14. [14]Shi Jin. Efficient asymptotic-preserving schemes for some multiscale kinetic equations. SIAM Journal on Scientific Computing, 21(2):441–454, 1999. doi: 10.1137/S1064827598334599. URL https://doi.org/10.1137/S1064827598334599.
  15. [15]Shi Jin, Zheng Ma, and Keke Wu. Asymptotic-preserving neural networks for multiscale time-dependent linear transport equations. arXiv preprint arXiv:2111.02541, 2021. doi: 10.48550/arXiv.2111.02541. URL https://arxiv.org/abs/2111.02541.
  16. [16]Shi Jin, Zheng Ma, and Keke Wu. Asymptotic-preserving neural networks for multiscale kinetic equations. arXiv preprint arXiv:2306.15381, 2023. doi: 10.48550/arXiv.2306.15381. URL https://arxiv.org/abs/2306.15381.
  17. [17]George E. Karniadakis, Ali Beskok, and Narayan Aluru. Microflows and Nanoflows: Fundamentals and Simulation. Springer, 2005. doi: 10.1007/0-387-28676-4. URL https://link.springer.com/book/10.1007/0-387-28676-4.DOI
  18. [18]Joonbeom Kim and Eunji Jun. Physics-informed neural networks based on the Fokker–Planck model for rarefied gas flows. arXiv preprint arXiv:2610.22390, 2026. doi: 10.48550/arXiv.2610.22390.DOI
  19. [19]Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. URL https://arxiv.org/abs/1412.6980.
  20. [20]Gyounghun Ko, Sung-Jun Son, Seung Yeon Cho, and Myeong-Su Lee. A theory-guided weighted L2 loss for solving the BGK model via physics-informed neural networks. arXiv preprint arXiv:2604.04971, 2026. doi: 10.48550/arXiv.2604.04971. URL https://arxiv.org/abs/2604.04971.
  21. [21]Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares. Quarterly of Applied Mathematics, 2(2):164–168, 1944. doi: 10.1090/qam/10666. URL https://doi.org/10.1090/qam/10666.
  22. [23]Liu Liu, Xueyu Zhu, and Zhenyi Zhu. A bi-fidelity based asymptotic-preserving neural network for the semiconductor boltzmann equation and its inverse problem. Journal of Computational Physics, 563:115076, 2026. doi: 10.1016/j.jcp.2026.115076. URL https://doi.org/10.1016/j.jcp.2026.115076.
  23. [24]Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://arxiv.org/abs/1711.05101.
  24. [25]Qin Lou, Xuhui Meng, and George Em Karniadakis. Physics-informed neural networks for solving forward and inverse flow problems via the Boltzmann–BGK formulation. Journal of Computational Physics, 447:110676, 2021. doi: 10.1016/j.jcp.2021.110676. URL https://doi.org/10.1016/j.jcp.2021.110676.
  25. [26]Donald W. Marquardt. An algorithm for least-squares estimation of nonlinear parameters. Journal of the Society for Industrial and Applied Mathematics, 11(2):431–441, 1963. doi: 10.1137/0111030. URL https://doi.org/10.1137/0111030.
  26. [27]Luc Mieussens. Discrete velocity model and implicit scheme for the BGK equation of rarefied gas dynamics. Mathematical Models and Methods in Applied Sciences, 10(8):1121–1149, 2000. doi: 10.1142/S0218202500000562. URL https://doi.org/10.1142/S0218202500000562.
  27. [28]H. M. Mott-Smith. The solution of the Boltzmann equation for a shock wave. Physical Review, 82(6):885–892, 1951. doi: 10.1103/PhysRev.82.885.DOI
  28. [29]Jaemin Oh, Seung Yeon Cho, Seok-Bae Yun, Eunbyung Park, and Youngjoon Hong. Separable physics-informed neural networks for solving the BGK model of the Boltzmann equation. SIAM Journal on Scientific Computing, 47(2):C451–C474, 2025. doi: 10.1137/24M1668809. URL https://doi.org/10.1137/24M1668809.
  29. [30]Maziar Raissi, Paris Perdikaris, and George E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019. doi: 10.1016/j.jcp.2018.10.045. URL https://www.sciencedirect.com/science/article/pii/S0021999118307125.
  30. [31]Ehsan Roohi and Christopher J. McDevitt. Regional accuracy of physics-informed neural networks (PINNs) for recirculation in a deep lid-driven cavity, 2026. URL https://arxiv.org/abs/2610.38814v2. Version 2, revised 31 October 2026.
  31. [32]Ehsan Roohi, Ahmad Shoja-Sani, and Fahimeh Ebrahimzadeh Aghzadi. Neural networks for rarefied gas dynamics: Relaxation problem, polyatomic shock waves, and hypersonic cylinder flow. Physics of Fluids, 38(5):057108, 2026a. doi: 10.1063/5.0334590. URL https://doi.org/10.1063/5.0334590.
  32. [33]Ehsan Roohi, Ahmad Shoja-Sani, and Stefan Stefanov. Physics constrained neural collision operators for hard sphere surrogates and ab initio angle prediction in direct simulation Monte Carlo. Physics of Fluids, 38(5):057123, 2026b. doi: 10.1063/5.0328463. URL https://doi.org/10.1063/5.0328463.
  33. [34]E. M. Shakhov. Generalization of the krook kinetic relaxation equation. Fluid Dynamics, 3:95–96, 1968. doi: 10.1007/BF01029546. URL https://doi.org/10.1007/BF01029546.
  34. [36]Yoshio Sone. Molecular Gas Dynamics: Theory, Techniques, and Applications. Birkhäuser, 2007. doi: 10.1007/978-0-8176-4573-1. URL https://link.springer.com/book/10.1007/978-0-8176-4573-1.DOI
  35. [37]Henning Struchtrup. Macroscopic Transport Equations for Rarefied Gas Flows. Springer, 2005. doi: 10.1007/3-540-32386-4. URL https://link.springer.com/book/10.1007/3-540-32386-4.
  36. [38]Henning Struchtrup and Manuel Torrilhon. Regularization of Grad’s 13 moment equations: Derivation and linear analysis. Physics of Fluids, 15(9):2668–2680, 2003. doi: 10.1063/1.1597472. URL https://doi.org/10.1063/1.1597472.
  37. [39]F. G. Tcheremissine. Conservative evaluation of Boltzmann collision integral in discrete ordinates approximation. Computers & Mathematics with Applications, 35(1–2):215–221, 1998. doi: 10.1016/S0898-1221(97)00269-1. URL https://doi.org/10.1016/S0898-1221(97)00269-1.
  38. [40]Manuel Torrilhon. Modeling nonequilibrium gas flow based on moment equations. Annual Review of Fluid Mechanics, 48:429–458, 2016. doi: 10.1146/annurev-fluid-122414-034259. URL https://doi.org/10.1146/annurev-fluid-122414-034259.
  39. [41]Manuel Torrilhon and Henning Struchtrup. Regularized 13-moment equations: shock structure calculations and comparison to Burnett models. Journal of Fluid Mechanics, 513:171–198, 2004. doi: 10.1017/S0022112004009917. URL https://doi.org/10.1017/S0022112004009917.
  40. [42]Jorge F. Urbán, Petros Stefanou, and José A. Pons. Unveiling the optimization process of physics informed neural networks: How accurate and competitive can PINNs be? Journal of Computational Physics, 523:113656, 2025. doi: 10.1016/j.jcp.2024.113656. URL https://doi.org/10.1016/j.jcp.2024.113656.
  41. [43]Tianbai Xiao and Martin Frank. Using neural networks to accelerate the solution of the Boltzmann equation. Journal of Computational Physics, 443:110521, 2021. doi: 10.1016/j.jcp.2021.110521.DOI
  42. [44]Tianbai Xiao and Martin Frank. RelaxNet: A structure-preserving neural network to approximate the Boltzmann collision operator. Journal of Computational Physics, 490:112317, 2023. doi: 10.1016/j.jcp.2023.112317.DOI
  43. [45]Kun Xu and Jian-Cheng Huang. A unified gas-kinetic scheme for continuum and rarefied flows. Journal of Computational Physics, 229(20):7747–7764, 2010. doi: 10.1016/j.jcp.2010.06.032. URL https://doi.org/10.1016/j.jcp.2010.06.032.
  44. [46]Linying Zhang, Wenjun Ma, Qin Lou, and Jun Zhang. Simulation of rarefied gas flows using physics-informed neural network combined with discrete velocity method. Physics of Fluids, 35(7):077124, 2023. doi: 10.1063/5.0156404.DOI
  45. [47]Pei Zhang and Yanli Wang. Neural network representation of microflows with BGK model. arXiv preprint arXiv:2410.21935, 2024. doi: 10.48550/arXiv.2410.21935.DOI

Paper details

vlnc:W_vw47d889
Version
v2 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

ChatGPT, Codex, Claude Fable 5.2

View statement
Looking for related papers…
Contents